Using machine learning to develop an asset allocation application

Abstract

This paper aims to analyze factor mimicking portfolio (FMP) construction in the context of Machine Learning and use it in a practical application. Inflation is chosen as the target macroeconomic factor in this study. We show that in the FMP construction process, Linear regression model performs better than LASSO and Ridge regression. In order to provide an application with FMP, Random Forest model is utilized in forecasting future inflation values so that asset allocation in the FMP can be done accordingly. While forecast accuracy of the Random Forest model is better than 1-period lag and simple moving average models, Rolling ARIMA model has smaller forecast error. However, FMP which is updated by the inflation forecasts from Random Forest provide more profit, higher Sharpe and Sortino ratios, and smaller maximum drawdown than the one with the Rolling ARIMA. Lastly, we follow a strategy such that we invest 95% of the initial capital to equally weighted portfolio and 5% of the initial capital to the FMP. We show that the proposed model outperforms the market where equally weighted portfolio of the same assets is used as market benchmark.